Comparison of Three Innovative Technologies for …Comparison of Three Innovative Technologies for...
Transcript of Comparison of Three Innovative Technologies for …Comparison of Three Innovative Technologies for...
Comparison of Three Innovative Technologies for
3D-Acquisition, Modelling, and Visualization of an Underground Mine
Valens FRANGEZ, Benjamin KRAMIS,
Florian HÜBSCHER, Andreas BAUMANN (Switzerland)
Key words: Terrestrial Laser Scanning, SLAM, Indoor Mobile Mapping, Underground Mine
Surveying, Virtual Reality, VR
SUMMARY
Mobile mapping (MM) solutions represent an innovative and effective approach for acquiring
geospatial information. While outdoor MM solutions are well established, mapping large-scale
indoor or underground environments with high surface complexity, detail and possibly low
ambient light is still a demanding task, and a variety of technological solutions start to emerge.
In this contribution, we compare three such solutions empirically with respect to data quality,
properties of the derived 3D models, and usability.
The analysis is based on data collected in a former gold mine using three different techniques.
A FARO Focus3D X330 was used to capture approximately 350 metres of underground
galleries employing terrestrial laser scanning. About 120 scans were taken within nine hours to
cover the entire space. Additionally, the same space was mapped independently using two
commercial MM systems suitable for indoor applications, namely a GeoSLAM ZEB-REVO
handheld scanner and a Leica Pegasus:Backpack MM system. Each of them required less than
20 minutes for acquiring the galleries in a single walkthrough. A geodetic terrestrial network
was established within the mine. Spherical targets placed in the scene and connected to the
network were used as a reference for the evaluation of the quality of the point clouds. A 3D
model was generated, fused with a digital terrain model and geological maps, and visualized in
a virtual reality environment using the HTC Vive headset.
We present the tools and procedures used for data acquisition, model derivation and
visualization including generation of fly-through videos. We also assess the solutions in terms
of quality, time consumption, complexity of the data processing, and potential use cases, giving
recommendations for tackling similar mapping and modelling tasks for underground or indoor
environments. None of the solutions stands out as the single optimum one. Rather, the choice
of the most appropriate solution depends on the intended application of the resulting 3d models.
Relevant questions are e.g., whether high accuracy, low cost for data acquisition or
photorealistic appearance are required. The MM solutions prove especially useful when a local
precision of a few cm is required but the absolute accuracy is of minor importance. Ground
control points throughout the mine could be used to improve the absolute accuracy. High time
consumption for both, scanning and preprocessing of the data (i.e. manual registration, etc.),
Comparison of Three Innovative Technologies for 3D-Acquisition, Modelling, and Visualisation of an Underground
Mine (9502)
Valens Frangez, Benjamin Kramis, Florian Hübscher and Andreas Baumann (Switzerland)
FIG Congress 2018
Embracing our smart world where the continents connect: enhancing the geospatial maturity of societies
Istanbul, Turkey, May 6–11, 2018
was recognized as the major downside of the terrestrial laser scanner as compared to the MM
solutions.
Comparison of Three Innovative Technologies for 3D-Acquisition, Modelling, and Visualisation of an Underground
Mine (9502)
Valens Frangez, Benjamin Kramis, Florian Hübscher and Andreas Baumann (Switzerland)
FIG Congress 2018
Embracing our smart world where the continents connect: enhancing the geospatial maturity of societies
Istanbul, Turkey, May 6–11, 2018
Comparison of Three Innovative Technologies for
3D-Acquisition, Modelling, and Visualization of an Underground Mine
Valens FRANGEZ, Benjamin KRAMIS,
Florian HÜBSCHER, Andreas BAUMANN (Switzerland)
1. INTRODUCTION
Mobile mapping (MM) solutions represent an innovative and effective approach for acquiring
geospatial information. Mapping large-scale indoor or underground environments with high
surface complexity, detail and possibly low ambient light is still a demanding task and a variety
of technological solutions start to emerge. Integrated geodetic solutions and systems addressing
these challenges are being developed for commercial purposes. The paper shows the advantages
and disadvantages of such solutions when scanning underground environments by comparing
the results to a standard approach employing a terrestrial laser scanner (TLS). We use a 350 m
long section of the former gold mine "La miniera d’oro di Sessa" (Switzerland) as a study case.
This is a particularly challenging underground environment because of the small tube diameter
(1.5 to 3 m).
The empirical comparison is based on data collected using three different instruments: a FARO
Focus3D X330 (F3D) terrestrial laser scanner, a Leica Pegasus:Backpack (LPB) MM system
and a GeoSLAM ZEB-REVO (ZEB) handheld MM system. We compare the point clouds
obtained using these instruments and the respective post-processing software addressing data
quality, properties of the derived 3D models, and usability of the instruments. The accuracy of
the point clouds is analysed using a geodetic network. Finally, we report about various
approaches to visualization of the point clouds including a virtual reality (VR) environment, a
fly-through video animation and publication of interactive 3D models on a website.
Similar topics have also been addressed by other authors in recent publications. They either
address challenges of underground scanning or present emerging MM solutions. (Sirmacek et
al., 2016) carried out a comparison of a terrestrial laser scanner and a MM system focusing on
their performance in the interior of a building. The challenge of mapping long and smooth
underground environments using a hand-held scanner is addressed in (Farella, 2016).
Furthermore, (Zlot et al., 2014) introduced an approach to underground mapping in the absence
of GPS by using a custom-built MM system. The past study cases show that the topic is relevant
and interesting for research from both, commercial and scientific point of view.
The paper is organized as follows. Section 2 presents the used methods including the
instruments and the acquisition process. Point cloud processing and geometric quality are
described in the section 3. Section 4 explains several different visualization approaches and
Comparison of Three Innovative Technologies for 3D-Acquisition, Modelling, and Visualisation of an Underground
Mine (9502)
Valens Frangez, Benjamin Kramis, Florian Hübscher and Andreas Baumann (Switzerland)
FIG Congress 2018
Embracing our smart world where the continents connect: enhancing the geospatial maturity of societies
Istanbul, Turkey, May 6–11, 2018
Section 5 presents the comparison of the results. Section 6 concludes the paper and gives
recommendations for future tasks.
2. FIELDWORKD AND PREPROCESSING
2.1 Instruments
The instruments used are shown in Figure 1. They differ in terms of handling, setup, data
acquisition process, and instrument specifications. Table 1 summarizes some instrument
specifications relevant for the carried-out project.
Figure 1: From left to right: FARO Focus3D X3301, Leica Pegasus:Backpack2, GeoSLAM ZEB-REVO3
According to the specification, the F3D TLS has the lowest ranging error of the instruments
used. It offers the possibility to modify the acquisition settings, namely data resolution and
quality, scan range, use of additional sensors during the scanning (e.g. compass, altimeter, GPS
or inclinometer) and a possibility to acquire colour information. Scanning with a lower
resolution allows faster capture of the environment and vice versa. During the scanning process,
the instrument must be mounted on a tripod. The acquired raw data are used to generate single
point clouds that are registered using the appropriate software (e.g. FARO Scene software).
The LPB is a wearable MM solution consisting of a backpack and a tablet for remote control
of the acquisition. The backpack is equipped with five cameras and two LiDAR profilers.
Additionally, data obtained with a GNSS antenna allow instant georeferencing of the acquired
data when satellite visibility and GNSS signal data quality are sufficient. An inertial
measurement unit (IMU) supports the georeferencing and aids in bridging GNSS data gaps. For
colouring the point clouds, spherical images are generated by stitching the images of all five
cameras together and projecting the RGB values onto the acquired point cloud. When mapping
dark areas an additional lighting device can be attached to the backpack. Additionally, the frame
rate of images and the camera exposure time can be modified. The acquired raw data are pre-
1 www.faro.com/en-gb/news/the-new-faro-laser-scanner-focus3d-x-330-the-perfect-instrument-for-3d-documentation-and-land-surveying-2 2 https://leica-geosystems.com/products/mobile-sensor-platforms/capture-platforms/leica-pegasus-backpack 3 https://geoslam.com/technology
Comparison of Three Innovative Technologies for 3D-Acquisition, Modelling, and Visualisation of an Underground
Mine (9502)
Valens Frangez, Benjamin Kramis, Florian Hübscher and Andreas Baumann (Switzerland)
FIG Congress 2018
Embracing our smart world where the continents connect: enhancing the geospatial maturity of societies
Istanbul, Turkey, May 6–11, 2018
processed and used to compute the 3D point clouds. The corresponding software uses a SLAM
algorithm combining laser scanning data with the GNSS/IMU data.
The ZEB is a handheld scanner that includes a 2D time-of-flight laser range scanner, which is
coupled to an IMU. No settings can be modified before the acquisition process. The operator
holds the scanner head while the data logger is carried along in a backpack. The environment
within the scanning range is acquired by a slow walking speed. Each scanning session is limited
to a maximum of 30 min. The acquired raw data are used to compute a point cloud using the
GeoSLAM Desktop software. The software uses a SLAM algorithm for creating a single 3D
point cloud.
Table 1: Instrument specifications according to the data sheets
F3D4 LPB5 ZEB6
Weight
(with batteries) 5.2 kg
13.7 kg
(backpack)
1.0 kg (scanning head)
4.1 kg (carry case)
Size 24 x 20 x 10 cm³ 73 x 27 x 31 cm³
8 x 11.3 x 14 cm³
(scanning head)
47 x 22 x 18 cm³
(carry case)
LiDAR type Proprietary 2 x Velodyne VLP-16 Hokuyo UTM-30LX
Ranging error 2 mm 20 – 30 mm 10 – 30 mm
Operation range
(distance measurement) 0.6 – 330 m 1 m – 100 m7 0.1 – 20 m
Acquisition up to 976'000 pts/sec 600'000 pts/sec 43'200 pts/sec
2.2 Data collection
2.2.1 Geodetic terrestrial network
A geodetic terrestrial network was established within the mine and at its entrance. It serves as
a reference for the quality assessment of the acquired point clouds. The network consists of two
parts, a portal network (8 points) and a subterraneous tunnel network (14 traverse points and 56
fixed points). The network measurements were carried out using a Leica TS60 total station. The
position of each fixed point had been planned such that it is visible from at least two traverse
points, the next and the previous one. Selected fixed points realised by bolts (holding prisms
during network measurements, and holding spheres later) served as the connection between the
network measurements and the acquired point clouds.
4 https://faro.app.box.com/s/4f908b59hcjjj8mezdr58z6n4qy5neli 5 https://leica-geosystems.com/-/media/files/leicageosystems/products/datasheets/leica_pegasusbackpack_ds.ashx?la=en 6 http://download.geoslam.com/docs/zeb-revo/ZEB-REVO%20User%20Guide%20V3.0.0.pdf 7 V.Marini, Geosoft srl / Leica Geosystems, personal communication (5.3.2018)
Comparison of Three Innovative Technologies for 3D-Acquisition, Modelling, and Visualisation of an Underground
Mine (9502)
Valens Frangez, Benjamin Kramis, Florian Hübscher and Andreas Baumann (Switzerland)
FIG Congress 2018
Embracing our smart world where the continents connect: enhancing the geospatial maturity of societies
Istanbul, Turkey, May 6–11, 2018
A minimum variance adjustment within a local coordinate system was carried out using the
portal network points as datum points. The standard deviations (1𝜎) as adjustment results for
the last network point in the mine are 5 mm, 14 mm and 1 mm in X, Y and H direction,
respectively.
2.2.2 Point cloud acquisition
The mine has a rough surface producing significant occlusions. Inside the mine, sparsely
distributed lamps provide low ambient lighting, leading to difficulties for acquiring surface
colour. The scanning workflow is different for the different instruments used, in particular
because MM relies on sensor motion during acquisition and requires input for bias/drift
mitigation (e.g. loop closures, short stops of the motion, etc.) while TLS needs to be static
during the measurement and thus requires careful selection of the instrument set-up locations
to obtain sufficient coverage. Some properties of the acquisition are gathered in the Table 2.
The spheres (diameters of 12 cm and 15 cm) were installed on selected points of the geodetic
network. With the standard TLS approach, they are used for registration. However, they also
mark reference points within the mine such that they can be identified and located within the
point clouds thus serving as a vital component for dataset comparison and quality analysis.
During the acquisition using MM, three spheres were placed at the entrance of the mine and
three at the far end of the mapped part of the mine. Ten more spheres were placed at locations
in between.
Since the F3D is stationary while scanning, occluded areas need to be identified and scanned
from additional stations. We used 117 stations to capture the rough surface in the long narrow
space inside the mine. Each scan took about five minutes. Due to the low scanner-to-surface
distances and in order to speed up the scanning process the ¼-resolution option (244’000
pts/sec) was selected and no images were taken using the internal camera. These images would
likely have been of limited use due to the lighting conditions. Registration was planned to be
carried out using targets instead of the point cloud itself; thus, at least three reference spheres
had to be covered commonly in any pair of point clouds.
The LPB allows to capture the whole mine in a single walkthrough at walking speed. The
acquisition using LPB started outside the mine in order to have the system initialized and the
data georeferenced based on the GNSS measurements. Performing the measurement in a closed
loop and placing the backpack at the initial position again at the end of the acquisition allows
the software to better correct for unavoidable time dependent drift even in absence of GNSS.
Having such loop closures or collecting data in double passes is a prerequisite to achieve the
accuracies specified in the data sheet. A further means to reduce drift effects is the use of zero-
velocity updates enabled by the operator stopping movement for a few seconds every few
minutes. While the LPB is a very versatile MM system, the specific mine was at the limit of its
applicability because (i) the acquisition process was physically demanding for the operator who
Comparison of Three Innovative Technologies for 3D-Acquisition, Modelling, and Visualisation of an Underground
Mine (9502)
Valens Frangez, Benjamin Kramis, Florian Hübscher and Andreas Baumann (Switzerland)
FIG Congress 2018
Embracing our smart world where the continents connect: enhancing the geospatial maturity of societies
Istanbul, Turkey, May 6–11, 2018
had to stoop constantly in order to avoid hitting the walls or ceiling with parts of the sensor
assembly within the narrow and low galleries, and (ii) the acquired surfaces were close to or
even below the minimum range of the system.
The acquisition process using the ZEB was done in a loop, ending approximately at the same
position as started. The sensor is swung while walking such that the point cloud produced by
the 2d-profile scanner covers the entire surfaces to be mapped. The walking speed can be
slowed down to increase the effective point density at areas where this is needed. In the present
study, this was done near the spheres at fixed points in order to assure that they are densely
covered.
Table 2: Key parameters of the acquisition
F3D LPB ZEB
Time [min] 540 10 16
Points [million] 5064 25 27
Raw data size [GB] 17 1 0.1
Radiometric data Intensity RGB /
3. POINT CLOUD PROCESSING AND GEOMETRIC QUALITY
3.1 Single-system processing
It is necessary to carry out a single-system processing (registration, data cleaning, subsampling,
etc.) on the point clouds before the acquired data are ready for comparison. The available F3D
is not able to register directly in the field; we chose to register the F3D scans using the spheres
and FARO Scene. The registration is calculated automatically once the centres of the spheres
are available, however identifying them required manually clicking on all the spheres, which
appear in the individual scans (in total approximately 600 clicks). The registered data can be
subsampled and exported into a common data format as single point clouds or complete point
cloud. We used the open source software CloudCompare for manual removal of the clutter,
which in this project resulted from mirroring effects in the puddles on the floor and other
unintentional features.
Leica carried out the collection and pre-processing of the LPB data. Georeferenced point clouds
with RGB colour, down-sampled to a nearly homogeneous point density of about 1/2.5 cm
using a voxel-based filter were provided for the present analysis. There are indications that the
glass of one scanner of the backpack may have been scratched. Additionally, due to the very
limited space inside the mine, the loop closure was not possible and the initialization before
entering the mine was impaired by poor GNSS availability. Therefore, the available dataset
represents only a one-way direction of acquisition and its accuracy and noise level do not
correspond to the potential of this MM system as specified in the data sheet. Unwanted features
Comparison of Three Innovative Technologies for 3D-Acquisition, Modelling, and Visualisation of an Underground
Mine (9502)
Valens Frangez, Benjamin Kramis, Florian Hübscher and Andreas Baumann (Switzerland)
FIG Congress 2018
Embracing our smart world where the continents connect: enhancing the geospatial maturity of societies
Istanbul, Turkey, May 6–11, 2018
(e.g. people, top part of the GNSS antenna) have been removed manually from the point cloud
for the further analysis.
The ZEB raw data were pre-processed using the corresponding software GeoSLAM Desktop
and exported in the selected data format. The fully automatic computing of the point cloud took
about 90 min on standard Windows PC, i.e. six times the time needed for data acquisition
(16 min). There was hardly any need for subsequent manual cleaning.
3.2 Initial quality analysis
Irrespective of any potential prior georeferencing we decided to transform all point clouds into
a common local coordinate system for further processing. This coordinate system is defined by
the centres of two spheres (baseline of 20 m) within the mine, but close to its entrance
(Figure 2). One point is fixing the position and the other one the rotation around z-axis. There
are no rotations around other axis applied, as the point clouds include the respective
instrument’s determination of the horizontal plane. By comparing the coordinates of the
remaining estimated sphere centres to the network coordinates the lateral, longitudinal (scale)
and height displacements are determined. As there is only one dataset acquired for each
instrument, no statement about repeatability is possible, and the results just represent indicative
snapshots. The calculated displacements visualised in Figure 3 reveal that the lateral deviations
obtained after georeferencing to the local coordinate system using only the two points near the
entry of the mine reach metre-level at the far end (inside the mine). Height and longitudinal
displacements are much lower while still showing significant drift. The results obtained using
the LTS (F3D) are significantly better as compared to the ground truth than the SLAM-based
techniques.
Figure 2: Aligned point clouds, using a 20-metre baseline, show significant discrepancies of a multiple of the mine’s diameter
(top view)
Comparison of Three Innovative Technologies for 3D-Acquisition, Modelling, and Visualisation of an Underground
Mine (9502)
Valens Frangez, Benjamin Kramis, Florian Hübscher and Andreas Baumann (Switzerland)
FIG Congress 2018
Embracing our smart world where the continents connect: enhancing the geospatial maturity of societies
Istanbul, Turkey, May 6–11, 2018
Figure 3: Lateral, longitudinal and height displacements between the point clouds and points of the terrestrial network.
Note the different scale for the lateral deviation
To check, to which degree the linear part of the observed coordinate drift, and the scale may be
due to inaccuracies of the sphere centres as detected within the point clouds the displacements
at the end of the mine are downscaled to the 20 m baseline length. The results are given in Table
3. The lateral displacements of 0.06 m (F3D), 1.00 m (LPB) and 0.69 m (ZEB) over 20 m are
on the order of one sphere radius for F3D and ten sphere radii for LPB and ZEB. This is far
more than what could be explained by potential inaccuracies of the sphere centre detection
within the point clouds, e.g. due to noise and limited point density. Thus, the deviations at the
end of the mine are a result of uncompensated drift rather than inaccurate georeferencing. This
indicates that ground control (or reference) points or other means of stabilizing the point clouds
are needed to obtain higher absolute accuracy.
Scale deviations and overall rotations can be addressed by applying a similarity transformation
using reference points at both ends of the mine. The results obtained after carrying out such a
transformation using three GCPs near the entrance and three near the end of the mine are also
shown in Table 3. The values represent the deviations from further reference points (not used
for the transformation) in the middle of the mine. They indicate that even after this rigid
transformation significant deviations still exist, especially for the lateral component of the LPB
(-3.82 m) and ZEB (-1.93 m) point cloud.
Comparison of Three Innovative Technologies for 3D-Acquisition, Modelling, and Visualisation of an Underground
Mine (9502)
Valens Frangez, Benjamin Kramis, Florian Hübscher and Andreas Baumann (Switzerland)
FIG Congress 2018
Embracing our smart world where the continents connect: enhancing the geospatial maturity of societies
Istanbul, Turkey, May 6–11, 2018
This analysis shows that the non-linear drift needs to be compensated for all measurement
systems used herein if cm-level accuracies are needed throughout the mine.
Table 3: Displacements (lateral, longitudinal and in height) of the point clouds with respect to the TPS network
(all values are given in metres)
a) At the end of the mine,
georeferenced using 2 GCPs
near entrance
b) Values from a) downscaled
to the 20 m baseline between
the 2 GCPs
c) At the center of the mine
after georeferencing using 3
GCPs at the entrance and 3 at
the end
Lat. Long. Height Lat. Long. Height Lat. Long. Height
F3D 1.02 -0.03 -0.02 0.06 0.00 -0.00 -0.20 -0.01 -0.01
LPB 16.21 -2.00 -1.94 1.00 -0.12 -0.12 -3.82 -0.94 0.19
ZEB 11.24 0.63 0.05 0.69 0.04 0.00 -1.93 -0.04 0.11
Figure 4: Differences between point clouds after rigid transformation using 3 GCP near the entry and 3 GCP near the far end
of the mine (version c) as of Table 3): F3D (green), LPB (red) and ZEB (violet).
3.3 Non-rigid transformation
Deformations of a point cloud – as resulting from sensor drift or accumulated registration
deviations in this study case – cannot be mitigated using a rigid-body or similarity
transformation. If the respective system allows using GCPs directly within the raw data
processing, this may mitigate the remaining deviations. If the system is or is operated as a black
box outputting registered/georeferenced point clouds like the MM in our study case, a different
approach is needed for drift mitigation. We propose to use non-rigid transformation based on a
sufficient number of GCPs distributed throughout the mine. Each of these GCPs needs to be
identified within the point cloud. The resulting displacement vectors between point cloud and
given GCP coordinates can then be spatially interpolated to obtain displacement vectors for all
points within the point cloud. Shifting each point by the corresponding displacement vector
then corresponds to a non-rigid transformation, which can reduce the deformations of the point
cloud. Typically, for preserving the high relative accuracy of neighbouring parts of the point
cloud, the chosen interpolation must assure that a closer GCP has a larger influence on the
transformation vector of a single point than a GCP at farther distance.
100 m
Comparison of Three Innovative Technologies for 3D-Acquisition, Modelling, and Visualisation of an Underground
Mine (9502)
Valens Frangez, Benjamin Kramis, Florian Hübscher and Andreas Baumann (Switzerland)
FIG Congress 2018
Embracing our smart world where the continents connect: enhancing the geospatial maturity of societies
Istanbul, Turkey, May 6–11, 2018
Unfortunately, no standard software for such non-rigid transformations was known or available
to the authors. Three established spatial interpolation schemes, namely nearest-neighbour,
triangle based interpolation and distance weighted interpolation, were therefore applied using a
proprietary implementation in Matlab. We found out that the nearest-neighbour interpolation
results in undesired discontinuities, and the triangle based approach fails in the present case
because of the geometric configuration: the GCPs are located almost along a line and very long,
narrow triangles result from the triangulation. However, the distance weighted interpolation
yielded a clear improvement in this specific case. Using all 16 GCPs marked by spheres during
the MM data acquisition the displacement vectors 𝒗𝑖 at the sphere centres were computed. For
each point of the cloud a displacement vector 𝒗(𝑿) was then calculated as a weighted sum of
the 𝒗𝑖vectors at the GCPs according to
𝒗(𝑿) = ∑𝒗𝑖
‖𝑿 − 𝑿𝑖‖2
16
𝑖=1
(1)
where the inverse of the squared distance between the interpolation point 𝑿 of the point cloud
and the respective GCP 𝑿𝒊 is used as weight. The squared distance in the weight factor was
chosen empirically as it produced the best result in this study case. The results obtained starting
from the LPB point cloud are shown in Figure 5 which displays the differences between the
F3D point cloud and the non-rigidly transformed LPB point cloud. The results obtained using
this simple non-rigid transformation are apparently already much better than those obtained
without taking the GCPs into account and help to bring the accuracy of the LPB results to about
the level expected if loop closures were available and the actual distances remained above the
minimum distances of the sensors.
Figure 5: Non-rigidly transformed LPB data (red) using the F3D sphere centres as GCP (green)
The results shown above indicate that it is worth considering non-rigid transformations for
improving underground MM point cloud accuracy. However, further investigations are needed
to (i) provide guidelines for the number and location of required GCPs for further reduction of
the deviations, and (ii) for selection of the most appropriate spatial interpolation function. It is
to be expected that the two questions are linked and need to be investigated jointly. These
investigations are left for future work.
100 m
Comparison of Three Innovative Technologies for 3D-Acquisition, Modelling, and Visualisation of an Underground
Mine (9502)
Valens Frangez, Benjamin Kramis, Florian Hübscher and Andreas Baumann (Switzerland)
FIG Congress 2018
Embracing our smart world where the continents connect: enhancing the geospatial maturity of societies
Istanbul, Turkey, May 6–11, 2018
4. VISUALIZATION
Visualization of geodetic data can be extremely useful when developing a project in
collaboration with people with or without any specific technical background, where it is
necessary to extract information out of the acquired data. Additionally, the project may have a
nontechnical purpose where data representation is of main interest. For this project, several
different innovative approaches of visualizing the acquired data were carried out, namely a
video animation, a 3D model print, and a virtual reality visualization. In addition, an interactive
website is used for publishing the project results and making them accessible to a wider
audience.
A mesh has to be generated out of the point cloud, as the first step for all the different
visualization approaches. The appropriate settings were chosen empirically, considering the
high point density of the original dataset, the desired resolution of the mesh and its visual
appearance, file size, etc. The imported raw point cloud has to be cleaned from artefacts and
later down-sampled. For the F3D data and the ZEB data, we have used a subsampling radius of
1.5 cm, in order to substantially reduce the number of points within the dataset and bring the
resolution to approximately the same level as the one that had been chosen by Leica for the
LPB data processing. Afterwards point normals were computed using quadratic local surface
models, which are suitable for local approximation of curvy and complex surfaces. We chose
to determine the orientation of the normals from samples of 40 points, respectively. After the
point clouds had been prepared in that way, the mesh was generated using the triangular mesh
generation algorithm (Poisson surface reconstruction). All these steps for the preparation of the
point clouds for visualization have been carried out using CloudCompare.
4.1 3D model print
The mesh of the gold mine represents the surface separating void space from solid
rock/surrounding material. For 3D printing, it has to be placed within a frame that supports the
whole structure (e.g. a box casing). To give the user access to the interior of the mine in this 3d
print, the box has to open up. We decided to realize this by splitting the entire model into an
upper and a lower part and allow the two parts to be folded together. The model for 3d printing
was therefore created using Boolean operations (e.g. subtraction) between two initial elements
– the mesh of the mine and the created box element. The walls of the created model have no
thickness, therefore a certain thickness needs to be added, depending on the printer
specifications, settings, and the used material (gypsum print requires a minimum thickness of
2 mm). We have additionally labelled the physical model by adding 3D text to the model. Figure
6 shows the printed parts of the model. The preparation of the data, starting from the generated
mesh takes a couple of hours, primarily depending on the file size of the mesh and the
performance of the used computer. Additionally, the duration of the 3D print depends on the
printer specification, settings (e.g. detail resolution) and the chosen material.
Comparison of Three Innovative Technologies for 3D-Acquisition, Modelling, and Visualisation of an Underground
Mine (9502)
Valens Frangez, Benjamin Kramis, Florian Hübscher and Andreas Baumann (Switzerland)
FIG Congress 2018
Embracing our smart world where the continents connect: enhancing the geospatial maturity of societies
Istanbul, Turkey, May 6–11, 2018
Figure 6: A picture of the printed 3D model of a cross intersection.
4.2 Video animation
A fly through animation of the point cloud data was created using the Bentley Pointools V8i.
Within the software, it is required to specify the poses and further parameters of the camera
along its desired path through the model. The effort and difficulty of creating the video
animation is low, however the time required to render the animation can take a few hours,
depending on the specified output settings (resolution, codec, video length, frames per second,
etc.). The final video obtained using the LPB point cloud, which was the only one in this study
containing RGB information, and thus the most suitable one for this type of visualization has
been published on YouTube8.
4.3 Virtual reality
We have also visualized the mesh in a virtual reality (VR) environment using the HTC Vive
headset. The visualization is achieved using the Unity game engine. We have designed three
levels of information (Figure 7) between which the viewer can freely change. The levels differ
in terms of information content and type of display.
The mesh representing the third level (c) was obtained by down-sampling the F3D point cloud,
since Unity does not perform well when using large data files (above 700 MB). Besides the
mesh, a point cloud is visualized and imported in Unity using the asset package Point Cloud
Viewer and Tools available on the Asset Store. The visualization of the model of the galleries
8 Gold Mine Visualization using Leica Pegasus:Backpack Data: https://www.youtube.com/watch?v=8VGN8Ccf2Rc&t
Comparison of Three Innovative Technologies for 3D-Acquisition, Modelling, and Visualisation of an Underground
Mine (9502)
Valens Frangez, Benjamin Kramis, Florian Hübscher and Andreas Baumann (Switzerland)
FIG Congress 2018
Embracing our smart world where the continents connect: enhancing the geospatial maturity of societies
Istanbul, Turkey, May 6–11, 2018
was fused with a digital terrain model (DTM), topographic and geological maps (together with
a legend of displayed rocks and minerals), as shown on the first (a) and second (b) level in
Figure 7. The DTM of Switzerland was obtained from swisstopo9 and the data extending over
the border to northern Italy were acquired from the European Environment Agency10 offering
open source data.
4.4 Online platform
Additionally, an interactive website is created for publishing the project results to the wider
audience using three individual approaches. The website is built up on Bootstrap, an open
source toolkit for developing with HTML, Cascading Style Sheets (CSS) and JavaScript.
The mesh generated from the F3D point cloud has been uploaded using Sketchfab. The platform
has a size limitation on the uploaded data of 50 MB. Therefore, the original point cloud was
down-sampled from 300 million to approximately 2 million points. This took around one hour
on a standard PC. Due to down-sampling the level of detail is of course reduced considerably,
but this is not a problem since the goal of the publication on the website is to convey an
impression about the mine, not to provide data for a technical analysis.
We have used Potree to visualize the coloured point cloud acquired by LPB. This is a web-
based renderer for large point clouds developed by Schütz (2016). The displayed point cloud
contains approximately 26 million points. Potree uses a hierarchical data structure with data
stored in sets of different resolutions. Because of this structure the whole point cloud can be
streamed and rendered within seconds. To create the corresponding rendered model took less
than halfe an hour on the PC mentioned above.
9 https://shop.swisstopo.admin.ch/en/products/height_models/dhm25 10 https://www.eea.europa.eu/data-and-maps/data/eu-dem
Figure 7: (a) First level of visualization using a topograpic map and a DTM; (b) second level of visualization using a
geological map and a DTM; (c) third level of visualization using a point cloud and a mesh
(a)
(b)
(c)
Comparison of Three Innovative Technologies for 3D-Acquisition, Modelling, and Visualisation of an Underground
Mine (9502)
Valens Frangez, Benjamin Kramis, Florian Hübscher and Andreas Baumann (Switzerland)
FIG Congress 2018
Embracing our smart world where the continents connect: enhancing the geospatial maturity of societies
Istanbul, Turkey, May 6–11, 2018
The data captured by the ZEB was visualized using Blend4Web Community Edition. This is an
open source platform allowing to publish computer graphics online. The point cloud data need
to be considerably down-sampled (in our case from approx. 6.5 million points to 800'000
points). On the other hand, the website then loads within a few seconds. The data are prepared
in the open source software Blender to obtain the desired appearance (surface colour,
reflectance, shadowing, etc.) and then exported into an HTML file ready for the upload11. To
prepare a scene using this approach still requires significant manual interaction and took about
a day in our case.
Sketchfab has the advantage that a mesh can be uploaded directly on the platform website, no
additional server is needed. This makes the handling easy and needs little time. The size
limitation on the uploaded data is a major restriction of this platform. Potree is the only platform
where the full point cloud can be uploaded without the need to down-sample the data. However,
no mesh visualization is possible. Blend4Web offers many adjustments to create the desired
appearance of the scene. Preparation is more time consuming for this software compared to the
other two platforms, and the mesh must be down-sampled significantly, because the browser
can otherwise not load the HTML file.
5. COMPARISON OF THE APPROACHES
5.1 Data quality and information content
The accuracy of the registered point clouds in terms of overall accuracy has been described
above (sec. 3). Using subsets of the data (indicated in
Figure 8) more specific aspects are briefly assessed in this subsection.
Figure 8: Locations of subsets used for detailed analysis
The spheres located within the mine play a crucial role in this paper, as they are used (i) to
establish a common coordinate system for all point clouds, and (ii) to assess the quality of the
registered point clouds. One of these spheres is shown in Figure 9 as represented within
different point clouds. They differ in amount, distribution and accuracy of the points due to the
different acquisition techniques, measurement processes and settings during both acquisition
and processing. The sphere is readily recognizable in the point cloud of the F3D because of the
high number of points (1828 pts) and low noise, largely due to the short distance between
scanner and sphere. The data provided by the LPB are comparatively sparse because they had
11 The website is available at: https://digitalreality.ethz.ch/goldmine
100 m
Fig. 9 Fig. 10
Fig. 11 Fig. 12
Comparison of Three Innovative Technologies for 3D-Acquisition, Modelling, and Visualisation of an Underground
Mine (9502)
Valens Frangez, Benjamin Kramis, Florian Hübscher and Andreas Baumann (Switzerland)
FIG Congress 2018
Embracing our smart world where the continents connect: enhancing the geospatial maturity of societies
Istanbul, Turkey, May 6–11, 2018
been reduced to a nearly constant point spacing of 2.5 cm for the entire mine by the Leica
operator. This yields a point cloud appropriately balancing detail versus file size for the entire
mine, but results in apparently few points (192) on the small spheres. While this impairs the
visual impression when just looking at the point clouds, and made the manual separation
between potential sphere points and other points (e.g. surrounding walls) harder, the number is
still sufficient to estimate the center of the sphere accurately for the analyses mentioned within
this paper. The ZEB data (2909 pts) show a very noisy, apparently egg-shaped point cloud
instead of a sphere. The shape is most likely the result of merging the data of forward and
backward acquisition fully automatically carried out by the respective software. The point
density at the sphere is apparently much higher than with the LPB although the full point clouds
of the mine have almost the same amount of data (see Table 2). The reason is that the ZEB
operator intentionally slowed down while walking past the spheres, moved closer and tried to
acquire the spheres with particularly many points, while the operation of the LPB required
maintaining nearly uniform walking speed and no special precautions were taken for dense
acquisition of the spheres.
Figure 9: Comparison of a sphere (radius 75 mm) as found in the point clouds avaialble for analysis
(from left to right: F3D, LPB, ZEB)
The point clouds obtained from the different techniques show very different distribution of
point densities over the whole mine (see Figure 10). As a stationary instrument with fixed
angular resolution, the F3D yields point densities decreasing linearly with distance. In the
present case, the highest density (~100’000 pts/m2) is found just above the instrument’s position
while the lowest density (~20’000 pts/m2) is found between the instruments and on the floor.
Due to the voxel-based filtering during processing, the LPB yields a point cloud with almost
constant point density (~6’000 pts/m2) except along two longitudinal sections of the ceiling
(~12’000 pts/m2). ZEB shows an irregular pattern of point density ranging from about 3’000
pts/m2 to about 20’000 pts/m2; this reflects the non-uniform movement of the observer and the
varying distances.
Figure 10: Point cloud density (different color map for the three images because of vastly different point densities; red: high
density, blue: low density).
F3D LPB ZEB
Comparison of Three Innovative Technologies for 3D-Acquisition, Modelling, and Visualisation of an Underground
Mine (9502)
Valens Frangez, Benjamin Kramis, Florian Hübscher and Andreas Baumann (Switzerland)
FIG Congress 2018
Embracing our smart world where the continents connect: enhancing the geospatial maturity of societies
Istanbul, Turkey, May 6–11, 2018
Cross-sections of the point clouds (see Figure 11) highlight the different noise levels, which are
mostly due to the different sensors involved and correspond to the respective specifications.
The static scanner (F3D) shows accordingly the highest precision. The LPB yields point clouds
with unevenly distributed noise exhibiting a much higher noise level in the upper half of the
cross-sections throughout the mine. We attribute this to the fact that the distance between the
sensors and the corresponding surfaces was less than 0.5 m and thus below the minimum
working range as specified by Leica. This suggests that data acquisition may be possible even
outside the specified range but at the cost of reduced precision. ZEB shows almost constant
noise level throughout the mine, but the point clouds are slightly noisier at areas which were
scanned with higher point density, e.g. close to the spheres.
Figure 11: Cross-sections of 5 cm width showing the noise levels of the point clouds and their spatial distribution at the
identical location within the mine.
Based on the specifications and the results of the above accuracy analysis, the F3D point cloud
seems to be the geometrically most accurate one. We have analysed deviations of the MM
results from this one also on a more local scale than above. A mesh-to-mesh (M2M) comparison
of an 8 m long section is shown in (Figure 12). The ZEB data mostly show small apparently
random deviations between -1 cm and 1 cm throughout the whole area, which is well within the
specification. The LPB results show systematic deviations exceeding 5 cm, again primarily in
the upper part of the gallery and probably linked to the operation outside the specified working
range. The M2M analysis was performed in GOM Inspect, while for all the other analyses
presented in this section CloudCompare was used.
Figure 12: M2M comparison with F3D mesh as reference
2 m
LPB ZEB F3D
LPB ZEB
Comparison of Three Innovative Technologies for 3D-Acquisition, Modelling, and Visualisation of an Underground
Mine (9502)
Valens Frangez, Benjamin Kramis, Florian Hübscher and Andreas Baumann (Switzerland)
FIG Congress 2018
Embracing our smart world where the continents connect: enhancing the geospatial maturity of societies
Istanbul, Turkey, May 6–11, 2018
While the LPB acquires RGB values for each point inherently, the F3D dataset shows only
reflectivity information in addition to the coordinates of the points. The ZEB provides neither
RGB nor reflectivity information. RGB information could have been acquired with the F3D at
the expense of approximately 50% increased scanning time. However, this would only have
worked appropriately if the galleries had been floodlighted sufficiently.
5.2 Technique comparison
The criteria to compare the deployed techniques and their results include the time of scanning
and processing, instrument handling, data complexity and accuracy, acquisition of additional
information (e.g. colour), etc. All the used instruments can be operated by a single person who
completed a system handling training. The LPB requires a basic training of 5-7 days for data
capture, processing and feature extraction, according to the manufacturer. The ZEB can be used
after an introduction of about an hour and by following the provided system manual, and the
training required for using the F3D and pre-processing its data to obtain a filtered and registered
point cloud is in the order of 1-2 days.
The F3D produced the best results in terms of bending, scale and lateral drift of the point clouds
as compared to the terrestrial network. The multiple acquisition settings allow using the
instrument according to the available time frame and required data quality. The major downside
of its use is the extended acquisition time. The resulting files are very large with vastly different
point densities depending on scanner-to-surface distances. Handling of the data is complex and
time-consuming if manual interaction is required (like for clicking onto the spheres and starting
manual registration in our case). Due to the rough surfaces, it might have been possible to do it
without spheres and register the point clouds directly using the iterative closest point (ICP)
algorithm (Besl and McKay (1992), Chen and Medioni (1991)). However, we would have
needed more overlap of the point clouds in that case and thus more scans. In our case spheres
at selected locations were used anyway for some of the analyses within this paper.
The advantage of LPB is the fast acquisition process, its mobility as compared to the F3D, and
the fact that useful RGB data are acquired along with the geometric data and without extra time
or effort. The instruments mobility, achieved by the backpack assembly, on the other hand
represents a limiting factor in very narrow spaces with low-ceiling like in this case due to the
backpack’s height when being worn. Actually, as stated above, the system had to operate below
its minimum specified distance within this mine. Targets or spheres placed in the scene are not
required. The generation of the final point cloud out of the raw data required approximately
four times the acquisition time length. A disadvantage in comparison with TLS is the lower
precision and the increased challenge to mitigate drift in case of unavailability of GNSS data.
The ZEB system, including the data processing part, has the advantage of being easy to use. In
particular, compared to F3D the acquisition time is very short. Because the scanner is handheld,
the user can move more freely than with a backpack system (like e.g. LPB). This was a clear
Comparison of Three Innovative Technologies for 3D-Acquisition, Modelling, and Visualisation of an Underground
Mine (9502)
Valens Frangez, Benjamin Kramis, Florian Hübscher and Andreas Baumann (Switzerland)
FIG Congress 2018
Embracing our smart world where the continents connect: enhancing the geospatial maturity of societies
Istanbul, Turkey, May 6–11, 2018
advantage in this narrow mine allowing also to increase point density at locations where this
was deemed useful (e.g. around spheres). The export and point cloud generation are fast and
require low effort. Targets or spheres placed in the scene are not required and the final point
cloud is generated automatically. Since no colour or intensity information is obtained, the data
are not as suitable for visualization purposes as the other data. There are very little opportunities
for the user to custom tailor parameters of data acquisition or data processing if needed.
6. CONCLUSIONS
The paper provides a comparison between three innovative technologies for 3D data
acquisition. Besides discussing the major advantages and disadvantages of the used
instruments, an empirical comparison is done. The comparison is based on various criteria: time
consumption, complexity of data processing, point cloud accuracy, etc. Since only one dataset
was acquired for each instrument, all the results are indicative only. A thorough accuracy
analysis based on a statistically valid sample would be highly interesting but is beyond the
scope of this paper. The paper involves new technological alternatives (LPB and ZEB) to well-
established terrestrial laser scanning (F3D) that are nowadays available on the market. Even
though the MM solutions have certain disadvantages, their use starts to address a wide range of
new application cases, which can also be of interest outside the surveying community.
Additionally, visualization of the geodetic data is becoming more and more important when
being used for touristic, architectural, archaeological, etc. purposes and therefore need to be
attractively presented. Several innovative approaches for visualizing the laser scanning data are
presented in the paper, making use of emerging technologies such as 3D print and VR.
The discussion of the results of the project illustrates the technical and practical advantages and
disadvantages of each technique. The pros and cons are summarized in Table 4, allowing the
user to identify whether the instrument will produce the most optimal results within certain
application.
Table 4: Summary of pros and cons of the techniques
F3D LPB ZEB
Pros
- Overall the best
performance when recording
complex surfaces over larger
scale, in terms of data quality
(low effects of bending, scale
and lateral drift),
- acquisition when high data
accuracy and resolution are of
importance,
- optional acquisition of RGB
and intensity information.
- Very fast acquisition without
much effort,
- targets/spheres not necessary in
the scene when scanning,
- acquisition when precision of a
few cm is required and the
absolute accuracy is of minor
importance,
- optional acquisition of RGB
information (allows further
visualization purposes).
- Very fast acquisition and
preprocessing of the data (point
cloud creation) and without
much effort,
- acquisition when precision of
a few cm is required and the
absolute accuracy is of minor
importance,
- handheld therefore more
flexible acquisition possible.
Comparison of Three Innovative Technologies for 3D-Acquisition, Modelling, and Visualisation of an Underground
Mine (9502)
Valens Frangez, Benjamin Kramis, Florian Hübscher and Andreas Baumann (Switzerland)
FIG Congress 2018
Embracing our smart world where the continents connect: enhancing the geospatial maturity of societies
Istanbul, Turkey, May 6–11, 2018
Cons
- High time consumption for
scanning (10 times longer as
compared to the other
techniques),
- targets/spheres necessary in
the scene when scanning,
- high data complexity and its
handling (e.g. manual
registration).
- Lower specified precision and
lower accuracy due to more
prominent drift,
- backpack’s height when worn
makes it hard to scan narrow and
low-ceiling environments,
- preprocessing of the data (point
cloud generation) is not straight
forward and long software
training is required.
- Lower specified precision and
lower accuracy due to more
prominent drift,
- little opportunity to custom
tailor parameters of data
acquisition,
- it does not provide RGB or
intensity information, therefore
the data cannot be used for
certain visualization purposes.
ACKNOWLEDGMENTS
We would like to thank everyone who was involved and helped to execute the project
throughout the semester. Especially Leica for providing the backpack, carrying out the data
acquisition and providing the processed data for free and Allnav for providing the ZEB-REVO.
E. Serantoni, Z. Gojcic (ETH) provided support throughout the project, the association ”Miniera
d’oro di Sessa” (A. Passera, G. Zanetti) assured support during data acquisition, A. Wieser
(ETH) gave feedback on an earlier version of the paper, and geosuisse awarded a grant enabling
participation at the conference.
REFERENCES
Besl P, McKay N (1992) A Method for Registration of 3-D Shapes. IEEE Trans. Pattern Anal.
Mach. Intell. 14:239–256.
Chen Y, Medioni G (1991) Object modeling by registration of multiple range images.
Proceedings 1991 IEEE Int Conf Robot Autom 2724–2729.
Farella, E. M. (2016). 3D Mapping of Underground Environments With a Hand-Held Laser
Scanner. In Proc.: 61° Convegno Nazionale SIFET, June 8-10, Lecce.
Schütz, M. (2016). Potree: Rendering Large Point Clouds in Web Browsers. Diploma thesis,
Vienna Institute of Computer Graphics and Algorithms, Vienna University of Technology.
Sirmacek, B. et al. (2016). Comparison of ZEB1 and Leica C10 Indoor Laser Scanning Point
Clouds. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information
Sciences, Volume III-1, pp. 143-149
Umeyama, S. (1991). Least-Squares Estimation of Transformation Parameters Between Two
Point Patterns. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 13, no.
4, pp. 376-380.
Wieser, A., Lienhart, W. and Brunner, F.K. (2003). Nachbarschaftstreue Transformation zur
Berücksichtigung von Spannungen im amtlichen Festpunktfeld. VGI – Österreichische
Zeitschrift für Vermessung und Geoinformation, vol. 91 (2), pp. 115-121.
Comparison of Three Innovative Technologies for 3D-Acquisition, Modelling, and Visualisation of an Underground
Mine (9502)
Valens Frangez, Benjamin Kramis, Florian Hübscher and Andreas Baumann (Switzerland)
FIG Congress 2018
Embracing our smart world where the continents connect: enhancing the geospatial maturity of societies
Istanbul, Turkey, May 6–11, 2018
Zlot, R. and Bosse, M. (2014). Efficient Large-Scale 3D Mobile Mapping and Surface
Reconstruction of an Underground Mine. In: Yoshida K., Tadokoro S. (eds) Field and Service
Robotics. Springer Tracts in Advanced Robotics, vol 92. Springer, Berlin, Heidelberg, pp. 479-
493.
BIOGRAPHICAL NOTES
Valens Frangez, Benjamin Kramis and Florian Hübscher are MSc students in Geomatic
Engineering at ETH Zürich. Andreas Baumann is a junior researcher at the Institute of Geodesy
and Photogrammetry, ETH Zürich.
CONTACTS
Andreas Baumann
Institute of Geodesy and Photogrammetry
ETH Zürich
Stefano-Franscini-Platz 5
8093 Zürich, Switzerland
Email: [email protected]
Comparison of Three Innovative Technologies for 3D-Acquisition, Modelling, and Visualisation of an Underground
Mine (9502)
Valens Frangez, Benjamin Kramis, Florian Hübscher and Andreas Baumann (Switzerland)
FIG Congress 2018
Embracing our smart world where the continents connect: enhancing the geospatial maturity of societies
Istanbul, Turkey, May 6–11, 2018